Is Using ML for Antivirus Safe

Sat Mar 23 2019

In this blog post I examine the ways in which antivirus programs currently employ machine learning and then go into the security vulnerabilities that it brings.

1 ML in the Antivirus Industry

Malware detection falls into two broad categories: static and dynamic analysis. Static analysis examines the program without actually running the code. Static analysis looks at things like the file fingerprints, hashes, reverse engineering, memory artifacts, packer detection, and debugging. Static analysis is largely known for looking up the hashes of the virus against a known database of viruses. It is super easy to fool signature based malware detection using simple obfuscation methods. Dynamic analysis is a technique where you run the program in a sandbox and monitor all the actions that the virus takes. If you notice that the program is acting suspicious, it is likely a virus. Suspicious behavior typically includes things like registry edits and API calls to bad host names.

Antivirus detection is very difficult, but, probably not for the reasons you think. The issue isn’t writing programs which can detect these static or dynamic properties of viruses– that is the easy part. It is also relatively easy to determine a general rule set for what makes a program dangerous. You can also easily blacklist suspicious domains, block malicious activity, and implement a signature based maleware detection program.

The real problem is that there are hundreds of thousands of malware applications and more are created every day. Not only are there tons of pesky malware applications, there is an absurd amount of normal programs which we don’t want malware applications to block. It is impossible for a small team of malware researchers to create a definitive set of heuristics which can correctly identify all malware programs. This is where we turn to the field of Machine Learning. Humans are very bad with big data, but, computers love big data. Most antivirus companies use machine learning and it has been a large success so far because it has allowed us to dramatically improve our ability to detect zero day viruses.

1.1 Interesting Examples

1.1.1 Cylance

Cylance uses supervised learning and static analysis to classify files as being malware. This product pulls a list of attributes from the file which they can then compare against other known viruses.

1.1.2 MalwareBytes Anomalous

Anomalous is a machine learning application which simply flags files which appear different from their training set of known normal files. This does not attempt to classify what makes a virus a virus, but, what makes a normal program a normal program. Anything which is not a normal program, it alerts you about since it can be a virus.

1.1.3 Kaspersky

Kaspersky appears to have done a ton of research into using machine learning for malware detection. I would highly recommend that you read their white paper on this subject.

2 Why is this a problem?

It turns out that machine learning systems can be easily fooled by using other machine learning algorithms. A classic example of this is with image classification. It is easy to use neural networks or genetic algorithms to generate examples which fool the machine learning application by learning the weights of the machine learning application and then making slight tweaks to your input to give a false classification.

Since viruses generation is a non-differentiable problem, people often use Genetic algorithms for the adversarial network to fool the antivirus. In other words, you don’t want to attempt to calculate the derivative between two versions of a virus for gradient decent. Since viruses are high dimensional problems, it turns out that most calc implementations would actually be inefficient at traversing the search space to find the global minimum. If you want to learn more about genetic algorithms, check out my recent blog post on it.

3 Fooling Antivirus Software

3.1 Genetic Algorithms

There are two major approaches which people have used to generate antivirus resistant malware with genetic algorithms. The first approach is to slowly make polymorphic changes to the virus in order to fool the malware detection. One of the interesting things about this approach is that you have to have some way of verifying that the polymorphic behaviors that you apply to the virus don’t break its “virus capabilities”.

An other approach used is to represent a virus as a set of properties. These properties are everything from the port of attack, the payloads, obfuscation parameters, etc. The genetic algorithm would simply tweak the properties of the virus until it found a configuration which evaded the antivirus program.

3.2 Reinforcement Learning

A research group at Endgame recently gave a Def Con talk where they presented a framework which uses reinforcement learning to evade static virus detection.

At a high level, the AI plays a “game” against the antivirus where the agent can make functionality-preserving mutations to the virus. The reward for the agent is its ability to not get detected by the anti-virus. Over time the AI will learn which type of actions will result in getting detected by the antivirus. This framework can be found on Github.

4 Takeaways

Machine learning is great, but, it needs to be properly defended. As we start to use machine learning more and more, a large portion of the cyber security field may shift its focus away from securing systems to securing big data applications.

5 Resources

• Blog post on Genetic Algorithms
• Kaspersky White Paper
• Windows Defender Use of ML
• Machine Learning Malware Models via Reinforcement Learning (Paper)
• Evolvable Malware (Paper)